Blame for research/src/arithmetic.py

"""Exact polynomial arithmetic, explicit cost vectors, and certified programs.

The field is F_65537. NumPy computes exact int64 schoolbook convolutions;
the configured size bound prevents integer overflow before modular reduction.
Field-operation counts model arithmetic, not machine instructions or time.
"""
from __future__ import annotations
from dataclasses import dataclass, asdict
from collections import Counter
import numpy as np
from synthesis import verify_tensor, linear_adds, digest

P = 65537
MAX_LENGTH = 8192

@dataclass
class Meter:
    add: int = 0
    mul: int = 0
    calls: int = 0
    plan_states: int = 0
    certificate_ops: int = 0
    allocations: int = 0
    def snapshot(self):
        d = asdict(self)
        d['field_ops'] = self.add + self.mul
        # Declared abstract ledger, not a measured instruction count.
        d['ledger_units'] = (self.add + self.mul + self.calls +
                             32*self.plan_states + self.certificate_ops)
        return d


def validate(a):
    a = np.asarray(a, dtype=np.int64)
    if a.ndim != 1 or len(a) < 1 or len(a) > MAX_LENGTH:
        raise ValueError('Polynomial lengths must be in [1, 8192]')
    if np.any(a < 0) or np.any(a >= P):
        raise ValueError('Coefficients must be canonical residues modulo 65537')
    return a


def base_cost(n, m):
    return n*m, n*m - (n+m-1)


class Kernel:
    """Interprets a finite, proof-carrying, depth-bounded arithmetic program.

    Every available program includes the schoolbook fallback. Selection is by
    exact dense field-operation cost, independent of coefficient values.
    """
    def __init__(self, law, cap=0):
        if type(cap) is not int or not 0 <= cap <= 20:
            raise ValueError('Invalid program depth')
        ok, work = verify_tensor(law)
        if not ok:
            raise ValueError('Rejected unsound bilinear tensor certificate')
        self.law = law
        self.cap = cap
        self.meter = Meter(certificate_ops=work)
        self.plans = {}
        self.trace = Counter()
        self.input_adds = sum(linear_adds(r) for r in law['A']+law['B'])
        self.output_adds = sum(linear_adds(r) for r in law['C'])

    def plan(self, n, m, cap=None):
        cap = self.cap if cap is None else cap
        key = (n, m, cap)
        if key in self.plans:
            return self.plans[key]
        self.meter.plan_states += 1
        bm, ba = base_cost(n, m)
        best = (bm, ba, 0)
        if cap > 0 and max(n, m) > 1 and min(n, m) > 1:
            h = (max(n, m)+1)//2
            cm, ca, _ = self.plan(h, h, cap-1)
            overhead = self.input_adds*h + (self.output_adds+2)*(2*h-1)
            candidate = (3*cm, 3*ca+overhead, 1)
            if candidate[0]+candidate[1] < bm+ba:
                best = candidate
            # A longer operand is split without permuting coefficient order.
            # This repairs the padding penalty on 2:1 Newton-lifting products.
            if max(n, m) >= 2*min(n, m):
                h = max(n, m)//2
                if n >= m:
                    x = self.plan(h, m, cap)
                    y = self.plan(n-h, m, cap)
                    extra = n-h+m-1
                else:
                    x = self.plan(n, h, cap)
                    y = self.plan(n, m-h, cap)
                    extra = n+m-h-1
                rectangular = (x[0]+y[0], x[1]+y[1]+extra, 2)
                if sum(rectangular[:2]) < sum(best[:2]):
                    best = rectangular
        self.plans[key] = best
        return best

    def cost(self, n, m):
        a, b, _ = self.plan(n, m)
        return a+b

    def _lin(self, arrays, row):
        nz = [(a, c) for a, c in zip(arrays, row) if c]
        if not nz:
            return np.zeros(len(arrays[0]), dtype=np.int64)
        a, c = nz[0]
        z = a.copy()
        self.meter.allocations += len(z)
        if c == -1:
            z = (-z) % P
            self.meter.add += len(z)
        for a, c in nz[1:]:
            z = (z + a) % P if c == 1 else (z - a) % P
            self.meter.add += len(z)
        return z

    def _mul(self, a, b, cap):
        n, m = len(a), len(b)
        self.meter.calls += 1
        _, _, split = self.plan(n, m, cap)
        if not split:
            cm, ca = base_cost(n, m)
            self.meter.mul += cm
            self.meter.add += ca
            self.meter.allocations += n+m-1
            # max coefficient sum <= 8192 * 65536**2 < 2**63.
            return np.convolve(a, b) % P
        if split == 2:
            h = max(n, m)//2
            if n >= m:
                left = self._mul(a[:h], b, cap)
                right = self._mul(a[h:], b, cap)
            else:
                left = self._mul(a, b[:h], cap)
                right = self._mul(a, b[h:], cap)
            out = np.zeros(n+m-1, dtype=np.int64)
            out[:len(left)] = left
            out[h:h+len(right)] = (out[h:h+len(right)]+right) % P
            self.meter.add += len(right)
            self.meter.allocations += len(out)
            return out
        h = (max(n, m)+1)//2
        aa = np.pad(a, (0, 2*h-n))
        bb = np.pad(b, (0, 2*h-m))
        av, bv = [aa[:h], aa[h:]], [bb[:h], bb[h:]]
        products = []
        for ar, br in zip(self.law['A'], self.law['B']):
            products.append(self._mul(self._lin(av, ar),
                                      self._lin(bv, br), cap-1))
        c = [self._lin(products, row) for row in self.law['C']]
        out = np.zeros(4*h-1, dtype=np.int64)
        out[:2*h-1] = c[0]
        for k in (1, 2):
            sl = slice(k*h, k*h+2*h-1)
            out[sl] = (out[sl]+c[k]) % P
            self.meter.add += 2*h-1
        self.meter.allocations += len(out)
        return out[:n+m-1]

    def mul(self, a, b):
        a, b = validate(a), validate(b)
        self.trace[(len(a), len(b))] += 1
        return self._mul(a, b, self.cap)

    def program(self, parent=None):
        body = {'schema': 1, 'kind': 'certified_bilinear_envelope',
                'field': P, 'cap': self.cap, 'law': self.law,
                'parent_sha256': parent, 'max_length': MAX_LENGTH}
        return {**body, 'sha256': digest(body)}


def load_program(data):
    if type(data) is not dict:
        raise ValueError('Invalid program envelope')
    body = {k: v for k, v in data.items() if k != 'sha256'}
    if digest(body) != data.get('sha256'):
        raise ValueError('Program integrity check failed')
    if body.get('field') != P or body.get('schema') != 1:
        raise ValueError('Unsupported program contract')
    if body.get('kind') != 'certified_bilinear_envelope':
        raise ValueError('Unsupported program language')
    return Kernel(body['law'], body['cap'])


class NTTKernel:
    """Conventional cached radix-two transform baseline, not a learned program.

    Roots, their construction, and modular inverse powers are charged. All
    transforms use exact field arithmetic. Schoolbook is selected whenever
    its predicted field-operation count is lower than a cached NTT estimate.
    """
    def __init__(self, law):
        self.meter = Meter()
        self.trace = Counter()
        self.roots = {}
        self.base = Kernel(law, 0)

    def _pow(self, a, k):
        y = 1
        while k:
            if k & 1:
                y = y*a % P
                self.meter.mul += 1
            k >>= 1
            if k:
                a = a*a % P
                self.meter.mul += 1
        return y

    def _roots(self, size, inverse):
        key = (size, inverse)
        if key in self.roots:
            return self.roots[key]
        w = self._pow(3, (P-1)//size)
        if inverse:
            w = self._pow(w, P-2)
        r = np.ones(size//2, dtype=np.int64)
        for j in range(1, len(r)):
            r[j] = r[j-1]*w % P
            self.meter.mul += 1
        self.roots[key] = r
        return r

    def _transform(self, a, inverse=False):
        n = len(a)
        bits = n.bit_length()-1
        # Integer indexing work is not included in field operation counts.
        rev = np.array([int(format(i, '0%db' % bits)[::-1], 2)
                        for i in range(n)], dtype=np.int64)
        out = a[rev].copy()
        size = 2
        while size <= n:
            r = self._roots(size, inverse)
            view = out.reshape(-1, size)
            left = view[:, :size//2].copy()
            right = view[:, size//2:]*r % P
            view[:, :size//2] = (left+right) % P
            view[:, size//2:] = (left-right) % P
            self.meter.mul += n//2
            self.meter.add += n
            size *= 2
        if inverse:
            inv = self._pow(n, P-2)
            out = out*inv % P
            self.meter.mul += n
        self.meter.allocations += n
        return out

    def mul(self, a, b):
        a, b = validate(a), validate(b)
        nout = len(a)+len(b)-1
        n = 1 << (nout-1).bit_length()
        if n > P-1:
            raise ValueError('Transform exceeds field root-of-unity contract')
        self.trace[(len(a), len(b))] += 1
        self.meter.calls += 1
        bm, ba = base_cost(len(a), len(b))
        estimate = (9*n*(n.bit_length()-1))//2 + 2*n
        if bm+ba <= estimate:
            self.meter.mul += bm
            self.meter.add += ba
            return np.convolve(a, b) % P
        aa = np.pad(a, (0, n-len(a)))
        bb = np.pad(b, (0, n-len(b)))
        c = self._transform(aa)*self._transform(bb) % P
        self.meter.mul += n
        return self._transform(c, True)[:nout]


def inv_scalar(x, backend):
    if not x % P:
        raise ValueError('Constant term is not a unit')
    k, a, out = P-2, int(x), 1
    while k:
        if k & 1:
            out = out*a % P
            backend.meter.mul += 1
        k >>= 1
        if k:
            a = a*a % P
            backend.meter.mul += 1
    return out